Skip to main content
BAN-02 · Banking & Capital Markets

Financial Crime Investigation Copilot

Outcome​

Compress alert-to-decision time for financial crime investigators by assembling transaction chains, KYC evidence, and watchlist matches into a narrative that an investigator confirms before any suspicious activity report is filed.

Portfolio attributeValue
Primary personaAML investigator
Primary aggregateInvestigationCase
Cosmos DB containers2
Grounding datatransactions, KYC, alerts, cases, watchlists
Development estimateUSD 60-220 monthly dev estimate
Runtime optionsPython FastAPI and C# ASP.NET Core

Productivity hypothesis​

These measures reflect the workflow, evidence, decision boundary, and industry risk for this use case.

Impact typeMeasureConservative internal assumptionCustomer measurement
ProductivityCase assembly cycle time8-15% reduction(baseline median case-assembly minutes - assisted median minutes) / baseline median minutes * 100
Quality or riskInvestigation rework5-10% reduction(baseline reopened or corrected investigation rate - assisted rate) / baseline rate * 100

Important: Conservative internal pilot assumptions only; not public benchmarks, commitments, or expected customer results. Validate against a comparable baseline or matched control cohort with mandatory human review.

Rationale: Transactions, KYC, alerts, cases, and watchlists support faster case assembly, not an unvalidated claim of improved crime detection.

Pilot guardrail: Preserve analyst judgment, evidence lineage, access controls, and formal case-handling procedures.

Customer validation plan​

  1. Baseline the current workflow with representative users and tasks.
  2. Run the accelerator on the same task mix with human review retained.
  3. Measure case assembly cycle time and investigation rework with the documented formulas.
  4. Replace internal assumptions with customer-observed results before a business case.

Why Azure Cosmos DB​

This AI solution needs an operational data plane, not only a vector index:

  • Hybrid vector and full-text retrieval combines semantic recall with exact identifiers, codes, names, and domain language.
  • Hierarchical partition keys and metadata predicates keep tenant and domain boundaries in the normal query path.
  • Operational JSON, source chunks, workflow and agent state, feedback, and audit records remain together for low-latency reads and governed writes.
  • Change feed reacts to source and operational updates without polling.
  • Autoscale throughput, configurable consistency, and multi-region distribution support variable demand and enterprise availability.

Domain workflow​

Domain workflow

StageActorInput → outputApproval
1. Triage and consolidate related alertsTriage agentAlert stream with subject identifiers → Consolidated case with linked alertsPolicy controlled
2. Resolve parties and beneficial ownershipEntity resolution serviceSubject identifiers and KYC records → PartyProfile set with ownership graphPolicy controlled
3. Reconstruct flow of fundsTransaction analytics agentTransaction ledger for the review window → TransactionChain with hop detail and totalsPolicy controlled
4. Draft the investigation narrativeNarrative agentChains, screening hits, and typology hypothesis → Cited NarrativeDraft with recommended dispositionPolicy controlled
5. Decide disposition and authorize filingInvestigator and compliance officerNarrative draft and evidence pack → Closed case or authorized regulatory filingHuman required

Technical architecture​

Technical architecture

Cosmos DB data model​

ContainerHierarchical partition keyDocumentsSearch paths
investigationCases/tenantId, /caseIdInvestigationCase, AlertRecord, TransactionChain, ScreeningHit, NarrativeDraft/adverseMediaVectorEmbedding, /narrativeVectorEmbedding, /adverseMediaVector, /narrativeVector
partyMasterProfiles/tenantId, /partyIdPartyProfile/profileNarrativeVectorEmbedding, /profileNarrativeVector

Query contract​

QueryContainerScopePurpose
Case evidence assemblyinvestigationCasessingle-partitionRead the full evidence set for one case in a single partition-aligned query.
Regulatory clock watchlistinvestigationCasescross-partition-boundedBounded cross-partition control-plane query: surface cases approaching the statutory reporting deadline. It cannot carry the second partition key because deadline risk is only visible when every open case partition is compared, so it runs as a tenant-scoped scheduled sweep off the interactive path, bounded by open cases whose regulatory due date falls inside a five-day horizon, executed with MaxItemCount paging and continuation tokens.
Unresolved screening hitsinvestigationCasescross-partition-boundedBounded cross-partition control-plane query: list high-score sanctions or adverse media matches still awaiting disposition across the case book. It cannot carry the second partition key because the sanctions queue is owned by a central team rather than a single case, so it runs as a tenant-scoped scheduled sweep off the interactive path, bounded by undecided dispositions above the analyst score threshold, executed with MaxItemCount paging and continuation tokens.
High-value structuring scaninvestigationCasescross-partition-boundedBounded cross-partition control-plane query: find chains whose aggregate value crosses the structuring review threshold. It cannot carry the second partition key because structuring is detected by comparing chains that belong to different cases, so it runs as a tenant-scoped scheduled sweep off the interactive path, bounded by a 30-day transaction window and a monetary review threshold, executed with MaxItemCount paging and continuation tokens.

Python and C# services route each document type to its declared container and hierarchical partition key. ETag-based writes protect mutable aggregates, while append-only records and bounded cross-partition queries preserve scale.

Azure deployment​

Azure deployment

Clone the source repository, authenticate with Azure, and run:

Set-Location accelerators/banking-capital-markets/financial-crime-investigation-copilot
az login
./Deploy.ps1 -Runtime python -Location eastus2

Change python to csharp to deploy the .NET implementation.

Explore the implementation​

Return to all accelerators